Optical modem fault prediction method based on timing analysis and GRU network

Abstract

This invention discloses a method for predicting optical modem faults based on time series analysis and a GRU network. This method first continuously collects basic operating status parameters, operating log data, and fault repair records from multiple enterprise-level optical modems. Key time series features, including the mean, variance, trend value, and number of alarm log entries of each type, are extracted from these data. Time windows are then constructed and labeled based on the fault repair records. Next, the corresponding time window vectors are input into a GRU network model for training. The trained GRU network model is then used to predict and alarm optical modem communication faults in real-world scenarios. This method enables the prediction of optical modem communication faults before they occur and preventive measures to be taken in advance, potentially avoiding failures and reducing communication losses. This method fills a gap in the domestic field of optical modem fault prediction technology.

Description

Technical Field

[0001] The present invention relates to the technical field of communication equipment fault prediction, and more particularly to an optical modem fault prediction method based on timing analysis and a GRU network. Background Art

[0002] Optical modems (also known as optical modems or single-port optical terminals) are essential communications devices for connecting fiber optic cables to businesses and organizations. With the increasing maturity of fiber optic technology, fiber-to-the-home (FTTH) has become the primary method for internet access, replacing traditional modems as the mainstream communications device for network access. However, due to the long-term operation of optical modems, which carry communication services such as fixed-line and broadband services, they sometimes experience communication failures such as dial-up disconnections and broadband connection drops. In severe cases, these failures can persist for days, causing significant communication losses for users. Predicting failures before they occur and taking preventive measures can, to a certain extent, avoid them and reduce communication losses. Therefore, research on communication failure prediction for optical modems is of great significance.

[0003] With the rapid development of machine learning in recent years, many researchers have attempted to apply various machine learning algorithms to fault prediction. Representative algorithms such as decision trees, logistic regression, k-means clustering, support vector machines (SVMs), and convolutional neural networks have been widely used in fault prediction for various devices, achieving significant results. However, the application of relevant models and algorithms in the field of fault prediction for communication equipment is still relatively limited. In particular, in the field of optical modem fault prediction, no relevant research has been conducted in China. Summary of the Invention

[0004] In response to the technical problems faced by the above-mentioned prior art, the present invention proposes an optical modem fault prediction method based on timing analysis and GRU network, which is used to predict the optical modem fault before it occurs, so that preventive measures can be taken in advance to avoid the occurrence of faults and reduce communication losses to a certain extent.

[0005] To achieve the above object, the present invention provides an optical modem fault prediction method based on timing analysis and GRU network, comprising the steps of:

[0006] Obtaining an original training data set of the optical modem including basic operating status parameters, operating log data, and fault repair records, and preprocessing the original training data set to obtain a purely numerical operating status parameter data set and an operating log data set;

[0007] Selecting an appropriate time window size and dividing the operating status parameter dataset and the operating log dataset into time windows to obtain a total dataset;

[0008] Performing feature extraction of state parameter information and log information on each time window in the total data set to obtain a time window vector and a time window stream;

[0009] According to the fault repair record, a label value is marked for each time window;

[0010] Taking the time window vector as input, building, training, and testing a GRU neural network model based on a training set and a test set divided by the time window stream according to a preset ratio to obtain a final-state GRU neural network model;

[0011] An optical modem fault prediction value is obtained based on the final-state GRU neural network model, and an optical modem fault is warned based on the prediction value.

[0012] Furthermore, the selecting of a suitable time window size and dividing the operating status parameter dataset and the operating log dataset into time windows to obtain a total dataset specifically includes:

[0013] Select the appropriate time window size based on the granularity requirements ;

[0014] Divide the operating status parameter data set into time windows and store consecutive timestamp values The state parameter vectors are divided into the same time window, and the operation state parameter vector array of the corresponding time window is obtained;

[0015] Divide the operation log data set into time windows, classify the log information whose occurrence time falls within a certain time window into the time window, and obtain the operation log vector array of the time window;

[0016] A total data set is obtained based on the operation status parameter vector array and the operation log vector array.

[0017] Further, .

[0018] Furthermore, the extracting features of state parameter information and log information from each time window in the total data set to obtain a time window vector and a time window stream specifically includes:

[0019] Perform feature extraction on the continuous value state parameter information, extracting one or more features including mean, variance, maximum value, minimum value and change trend to obtain the first feature vector of each continuous value state parameter in the current time window;

[0020] Performing feature extraction on the discrete value state parameter information, extracting one or more features including an initial stable value, a number of changes, and a final stable value, to obtain a second feature vector of each discrete value state parameter in the current time window;

[0021] Perform feature extraction on the log information, traversing the configuration information sequence, alarm information sequence, and error information sequence of each time window in turn, counting the number of log information of various configuration types, alarm types, and error types to obtain the third feature vector of the log information in the current time window;

[0022] The time window vector and the time window stream are acquired based on the first eigenvector, the second eigenvector, and the third eigenvector.

[0023] Furthermore, extracting the characteristics of the change trend specifically includes:

[0024] Defining monotonic coefficients , used to characterize the monotonicity of the continuous-valued state parameter within the current time window;

[0025] Define the amplification factor , used to characterize the magnitude of the change in the continuous value state parameter within the current time window;

[0026] Changing trends The calculation formula is:

[0027] ;

[0028] in, and is the weight coefficient, , , the weight coefficient is used to adjust and The relative weight of the impact on the changing trend.

[0029] Furthermore, the step of labeling each time window according to the fault repair record specifically includes:

[0030] According to the fault repair record, all time points at which communication faults occurred are obtained, and the time window label value to which these time points belong is set to 1, indicating that the time window is the fault window in which the communication fault occurred;

[0031] Defining constants , for the continuous time window, and set its label value to ,in , that is, the current window With the fault window The difference between the time points indicates the distance between the current window and the fault window;

[0032] For all distances to the fault window The time windows are all considered normal windows and their label values are set to 0.

[0033] Furthermore, the constructed GRU neural network model specifically includes an input layer, a hidden layer, and an output layer;

[0034] Input layer: The input parameter is the time window vector;

[0035] Hidden layer: Use a single hidden layer structure; record the current input as , the reset gate and update gate are respectively denoted as and , the hidden state of the previous moment is recorded as , the hidden state at the current moment is recorded as , the candidate hidden state at the current moment is recorded as , then reset the gate and update gate The calculation formula is as follows:

[0036] ,

[0037] ,

[0038] in, 、 、 and is the weight parameter, and is the bias parameter;

[0039] Candidate hidden states The calculation formula is as follows:

[0040]

[0041] in, and is the weight parameter, is the bias parameter, symbol yes Product operator; use A nonlinear activation function is used to ensure that the value of the candidate hidden state remains in the interval (-1, 1). Its expression is:

[0042]

[0043] at last, Hidden state at the moment The calculation formula is as follows:

[0044]

[0045] Output layer: output predicted value , represents the probability of failure in the current operating state, and its expression is:

[0046]

[0047] in, is the output layer weight matrix, is the output layer bias parameter, using as an activation function.

[0048] Furthermore, Huber's Robust Loss is used as the loss function during training, and its expression is:

[0049]

[0050] in, Indicates the actual label value, Represents the predicted value of the model output;

[0051] Use the mini-batch stochastic gradient descent algorithm to get the overall loss by taking the average error loss.

[0052] Furthermore, training and testing the GRU neural network model based on the training set and test set divided according to a preset ratio of the time window stream specifically includes:

[0053] The holdout method is used to divide the time window stream that has completed feature extraction into A continuous subsequence, and then The molecular sequences are divided into training set and test set according to the preset ratio, while keeping the proportion of subsequences containing fault windows in the training set and the test set equal to the total subsequences;

[0054] The training set is input into the GRU neural network model, the learning rate is set to 0.03~0.1, the number of training iterations is 1000, and the model training is started; after the training is completed, the test set is input into the GRU neural network model, the test is performed to obtain the final loss value, and the learning rate with the minimum loss value is used to train the model again to obtain the final state GRU neural network model.

[0055] Furthermore, the obtaining of an optical modem fault prediction value based on the final state GRU neural network model and issuing an early warning of the optical modem fault based on the prediction value specifically includes:

[0056] The operating status parameters are read once every first preset time, and the operating log data is read once every second preset time. The status parameters and log data are merged, the time window features are extracted, and the time window is constructed. The time window vector is input into the final state GRU neural network model to obtain the output value. When the output prediction value is greater than the preset threshold, it is judged that a fault is about to occur and a fault alarm is issued. Otherwise, no response is given. At the same time, the hidden state of this time window is stored and participates in the hidden state calculation at the next time point, and this cycle is repeated.

[0057] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0058] The method of the present invention first continuously collects basic operating status parameters, operation log data, and troubleshooting records from multiple enterprise-level optical modems. Key time series features, including parameter means, variances, trend values, and the number of alarm log entries of various types, are extracted from these data. Time windows are then constructed and labeled based on the troubleshooting records. Next, the corresponding time window vectors are input into a GRU network model for training. The trained GRU network model is then used to predict and issue alarms for optical modem communication failures in real-world scenarios. This allows for the prediction of optical modem communication failures before they occur and for preventive measures to be taken in advance, potentially avoiding failures and reducing communication losses. This method fills a gap in the domestic field of optical modem fault prediction technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 A schematic diagram of the overall process of an optical modem fault prediction method based on timing analysis and GRU neural network provided by an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of the data structure after time window division provided by an embodiment of the present invention;

[0062] Figure 3 A schematic diagram of the data structure after window feature extraction provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0064] The terms "first," "second," or "third" in the specification, claims, or drawings of this application are used to distinguish between different objects, not to describe a specific order. In addition, the terms "including" or "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to the process, method, product, or apparatus.

[0065] like Figure 1 As shown, in one embodiment, a method for predicting optical modem faults based on timing analysis and GRU neural network mainly includes steps S1-S6.

[0066] S1. Obtain an original training data set of the optical modem including basic operating status parameters, operating log data, and fault repair records, and pre-process the original training data set to obtain a purely numerical operating status parameter data set and an operating log data set. More specifically, the process includes steps S11-S12.

[0067] S11. Through the intelligent network management system, the basic operating status parameters, operation log data and fault repair records of multiple gigabit enterprise-level optical modems are continuously and uninterruptedly collected as the original training data set.

[0068] Basic operating status parameters mainly include:

[0069] 1) Optical module information: such as light status, transmit optical power, receive optical power, bit error rate and operating temperature, etc., recorded as ;

[0070] 2) Core module information: such as CPU usage, memory usage, device operating time, CPU temperature, motherboard temperature, operating voltage, operating current and total number of connection ports, etc., recorded as ;

[0071] 3) WLAN interface status information: such as the number of bytes received / sent, the number of frames received / sent, the number of error frames received / sent, the number of discarded frames received / sent, etc., recorded as ;

[0072] 4) LAN1~LAN8 interface status information: such as each port connection status (online / offline), connection mode (full duplex / half duplex), connection rate, number of bytes sent / received per port, number of frames sent / received per port and number of error frames, etc., recorded as .

[0073] The sum of these parameters is recorded as the state parameter information of the optical modem, and a vector is used Indicates that:

[0074]

[0075] in It is the timestamp, that is, the time point of collection; That is to say The optical modem state parameter vector at the moment;

[0076] use Represents all the collected state parameter data sets, then can be represented as an array of state parameter vectors arranged in time order; that is:

[0077] ( is a time series number)

[0078] The operation log data records the operations and exceptions that occur during the operation of the optical modem. The information in the operation log can be divided into three types: configuration information, alarm information and error information. The content of the configuration information includes the operation time and operation type ; The content of the alarm information includes the alarm time and alarm type ; The content of the error message includes the reporting time and error types .

[0079] use express A piece of log data generated at each moment, as described above, It can be a configuration log data, an alarm log data or an error log data. Depending on the type, it can be represented as follows:

[0080] or or

[0081] use represents the total running log data set, then It can be represented as an array of running log information arranged in chronological order, namely:

[0082]

[0083] Note that the subscript here 、 and It is not necessarily a continuous natural number, because the time when the log is generated is not fixed and discontinuous, unlike the status parameter information, which is always generated at every moment.

[0084] S12. Perform necessary data cleaning and data preprocessing on the data, including converting character data into numerical data, converting date data into timestamp, and normalizing the data. After processing, a purely numerical operating status parameter data set is obtained. and run log datasets .

[0085] S2. Select an appropriate time window size and divide the operating status parameter dataset and the operating log dataset into time windows to obtain a total dataset. To extract temporal features from the data and reduce the data size, the data needs to be further divided into time windows. More specifically, steps S21-S23 are included.

[0086] S21. First, select the appropriate time window size based on the actual scenario and granularity requirements. .

[0087] According to experience, Should meet ;like If the value is too small, it will not be possible to effectively reduce the data size and extract the temporal characteristics of the data, resulting in a long training time. If the value is too large, the number of divided time windows will be too small, resulting in underfitting of the training effect. as the time window size.

[0088] S22, the next step is to analyze the running status parameter data set Processing: The timestamp value is continuous state parameter vectors are divided into the same time window, that is, Divided into:

[0089]

[0090] Among them, the shape is a time window, denoted as ,Right now:

[0091] .

[0092] S23, then the operation log data set Processing: The log data appearing in each time window is classified into the corresponding time window as one of the features of the time window.

[0093] The specific operation is: , in the dataset Search for timestamps falling within Log information within the interval is added to middle.

[0094] After this step, a time window can be expressed as:

[0095]

[0096] in, Represents the operating status parameter vector array of the time window, 、 and They represent the configuration information sequence, alarm information sequence, and error information sequence within the time window, that is, the operation log vector array of the time window.

[0097] After this step, the total data set can be expressed as:

[0098] .

[0099] The data structure diagram after time window division is as follows Figure 2 shown.

[0100] S3. Perform feature extraction of state parameter information and log information on each time window in the total data set to obtain a time window vector and a time window stream. More specifically, the process includes steps S31-S32.

[0101] S31. Feature extraction is performed on the state parameter information. State parameters are classified into continuous values and discrete values. Continuous values refer to state parameters that continuously change during operation, such as CPU temperature, memory usage, operating voltage, and transmitted optical power. Discrete values refer to state parameters that have a single value and generally do not change, such as the light state of an optical module, the connection mode and connection status of a port, etc.

[0102] S311. Extract features of continuous value state parameter information.

[0103] For each continuous valued state parameter , extract the following 5 features: mean ,variance , maximum value , minimum value and changing trends .

[0104] 1) Mean Calculation formula:

[0105]

[0106] 2) Variance Calculation formula:

[0107]

[0108] 3) Maximum value and minimum value The calculation formula is:

[0109] ,

[0110] 4) Trends It is a special parameter defined in this invention, used to characterize the changing trend of the variable within the time window; , which is calculated as follows:

[0111] 41) Define the monotonic coefficient , used to characterize the state parameters Monotonicity within the current time window:

[0112]

[0113] in, The expression is:

[0114]

[0115] Under this definition, The larger the absolute value of the state parameter The stronger the monotonicity within the time window; When , the overall trend is upward; when , the overall trend is downward.

[0116] 42) Define the amplification factor , used to characterize the state parameters The magnitude of the change within the current time window:

[0117]

[0118] Under this definition, The larger the absolute value of the state parameter The greater the change in the current time window, the greater the hour, The value of is positive if it increases within the current window; otherwise, it is negative.

[0119] 43) Changing trends The calculation formula is:

[0120]

[0121] in, and is the weight coefficient , used to adjust and The relative weight of the impact on the changing trend.

[0122] Under this definition, when When the state parameter In the current time window, the overall trend is increasing; when hour, In the current window, the overall trend is decreasing; and, The larger the absolute value of , the more significant the growth / decrease trend.

[0123] After this step of extraction, each continuous value state parameter Feature availability vector within the current time window Expressed as:

[0124] .

[0125] S312: Extract features from discrete value state parameter information.

[0126] For each discrete value state parameter , extract the following three features: the initial stable value (within the current time window) , number of changes and the final stable value .

[0127] Among them, the initial stable value refers to the state parameter The value at the starting time point of the current time window; the last stable value refers to the state parameter The value at the end time of the current time window; the number of changes refers to the state parameter The number of times the value changes in the current time window. The first two eigenvalues are obtained by directly collecting the start and end values of the time window, and the number of changes is obtained by traversing the entire time window data once and counting The number of times the value of is changed is obtained.

[0128] In most time windows, the number of changes is 0, and the initial stable value and the final stable value Same; in a few time windows, these values have one jump (such as the port status changes from connected to disconnected), that is, the number of changes is 1. At this time, ; In very rare time windows, these values jump repeatedly (such as repeated power failure and reset), that is, the number of changes is greater than 1. In the above cases, 、 and All three parameters can effectively describe the changing characteristics of the parameter in this time window.

[0129] After this step of extraction, the feature vector of each discrete value state parameter in the current time window is available Expressed as:

[0130] .

[0131] S32, extract features from log information. For each time window , traverse its configuration information sequence in turn , Alarm information sequence and error message sequence ,Count the number of log messages of various configuration types, alarm types and error types as the ,features of the time window.

[0132] Feature vector of log information after extraction Expressed as:

[0133] .

[0134] in, Indicates the The number of log entries for each configuration type. Indicates the Number of log entries of each alarm type, Indicates the Number of log entries of each error type.

[0135] At this point, the features of the time window have been extracted and can be used as the time window vector Expressed as:

[0136] .

[0137] in, The characteristic vectors of each continuous-valued state parameter are represented in turn (as mentioned above, including mean, variance, maximum, minimum and change trend); The characteristic vectors of each discrete value state parameter (including the initial stable value, the final stable value and the number of changes) are represented in turn. The feature vector representing the log information. These parameters will be used as input to train the GRU neural network. The final time window stream is:

[0138] .

[0139] The data structure diagram after window feature extraction is as follows: Figure 3 shown.

[0140] S4. According to the fault repair record, a label value is added to each time window. More specifically, the process includes steps S41-S43.

[0141] S41. First, according to the fault repair record, obtain the time points of all communication faults, and set the time window label value of these time points to , indicating that the time window is a fault window (a fault occurred within the time window).

[0142] S42. Secondly, it is reasonable to infer that the time window close to the fault window has a certain degree of fault occurrence rate. Therefore, a constant is defined , for the continuous window, set its label value to ,in The current window With the fault window The difference between the time points of the fault window and the fault window indicates the distance between this window and the fault window.

[0143] S43. For all distances to the fault window The time window is considered to be a normal window, and its label value is set to .

[0144] S5. Using the time window vector as input, and based on the training set and test set divided by the time window stream according to a preset ratio, construct, train, and test a GRU neural network model to obtain a final-state GRU neural network model. More specifically, the process includes steps S51-S52.

[0145] S51, the specific construction of the GRU network model. The GRU network model, that is, the Gated Recurrent Unit (GRU) neural network, is a variant of the recurrent neural network (RNN) and is widely used in the processing of time series models. The key difference between the gated recurrent unit neural network and the ordinary recurrent neural network is that it supports the gating of hidden states. Specifically, it adds a reset gate and an update gate on the basis of the ordinary recurrent neural network. Among them, the reset gate helps to capture the short-term dependencies in the sequence, and the update gate helps to capture the long-term dependencies in the sequence. Due to the addition of these two gating units, GRU has a stronger ability to capture key information than ordinary recurrent neural networks. The present invention performs training and prediction of a communication fault prediction model based on the GRU neural network, and the structure of the GRU neural network is described as follows.

[0146] 1) Input layer: The input parameter is the time window The characteristic vector of (that is, the time window vector), that is ; In the following, the input is denoted as .

[0147] 2) Hidden layer: A single hidden layer structure is used, the number of hidden layer neurons is set to 512, and the input layer is fully connected to the hidden layer; the reset gate and update gate are also set to a 512×1 matrix.

[0148] Record the current input as , the reset gate and update gate are respectively denoted as and , the hidden state of the previous moment is recorded as , the hidden state at the current moment is recorded as , the candidate hidden state at the current moment is recorded as , then reset the gate and update gate The calculation formula is as follows:

[0149] ,

[0150] ,

[0151] in, 、 、 and is the weight parameter, and is the bias parameter;

[0152] Candidate hidden states The calculation formula is as follows:

[0153] ,

[0154] in, and is the weight parameter, is the bias term (bias parameter), symbol yes Product (element-wise product) operator. Here, we use A nonlinear activation function is used to ensure that the value of the candidate hidden state remains between (-1, 1), and its expression is:

[0155]

[0156] at last, Hidden state at the moment The calculation formula is as follows:

[0157] .

[0158] 3) Output layer: Output is a predicted value , which indicates the probability of failure in the current operating state:

[0159] ,

[0160] in, is the output layer weight matrix, is the output layer bias term. as an activation function.

[0161] Loss function: In order to improve the robustness of the model and prevent the problem of gradient explosion, it is preferred to use Huber's Robust Loss as the loss function during training, and its expression is:

[0162]

[0163] in, Indicates the actual label value, Represents the predicted value of the model output. This loss function effectively alleviates the problem of gradient explosion by reducing the loss function to linear when the difference is greater than 1.

[0164] Optimization algorithm: Use small batch stochastic gradient descent algorithm. The traditional stochastic gradient descent method needs to calculate the error loss on the entire training set each time, which is too expensive. The present invention preferably uses the small batch stochastic gradient descent algorithm. The specific method is: each time select time window, and calculate the difference between its predicted value and the true label , find the average error loss, to approximate the equivalent of the overall loss, that is:

[0165]

[0166] Initialization of the model: All weight matrices are randomly initialized to random number matrices with a mean of 0 and a variance of 0.1; all bias parameters are initialized to a 0 matrix; at the same time, in order to adapt to the situation where the first window in the time series has no preceding hidden state, the hidden state is initialized to a 0 matrix before training.

[0167] S52, GRU neural network training and testing. Using the holdout method, the time window stream (i.e. ) is divided into A continuous subsequence, and then The molecular sequences are divided into training set and test set in a ratio of 9:1, while keeping the proportion of subsequences containing fault windows in the training set and test set equal to the total subsequences.

[0168] The training set is input into the model, and the learning rate is set to 0.03~0.1 in sequence, the number of training iterations is 1000, and the model training is started; after the model training is completed, the test set is input into the model, and the final loss value is obtained after testing. The learning rate with the minimum loss value is used to train the model again to obtain the final model, that is, the final state GRU neural network model.

[0169] S6. Obtain an optical modem failure prediction value based on the final-state GRU neural network model, and issue an early warning of an optical modem failure based on the prediction value. More specifically, during the actual operation of an enterprise-level optical modem, the operating status parameters are read every 30 seconds, and the operating log data is read every 30 minutes. Then, as described above, the status parameters and log data are combined, and time window features are extracted to construct a time window. The corresponding time window vector is input into the model to obtain an output value. When the output prediction value is greater than 0.6, it is considered that a failure is about to occur, and a fault alarm is issued; otherwise, no response is given. Simultaneously, the hidden state of this time window is stored and used in the hidden state calculation at the next time point, and this cycle repeats.

[0170] It should be noted that the flowcharts or block diagrams in the accompanying drawings illustrate the possible architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a portion of code, and the above-mentioned module, program segment, or a portion of code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified functions or operations, or can be implemented with a combination of dedicated hardware and computer instructions.

[0171] Those skilled in the art will appreciate that the features described in the various embodiments and / or claims of the present disclosure may be combined and / or coupled in various ways, even if such combinations and / or couplings are not explicitly described in the present disclosure. In particular, the features described in the various embodiments and / or claims of the present disclosure may be combined and / or coupled in various ways without departing from the spirit and teachings of the present disclosure, and all such combinations and / or couplings fall within the scope of the present disclosure.

[0172] Although the present disclosure has been shown and described with reference to certain exemplary embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made to the present disclosure without departing from the spirit and scope of the present disclosure as defined by the appended claims and their equivalents. Therefore, the scope of the present disclosure should not be limited to the above-described embodiments, but should be determined not only by the appended claims but also by the equivalents of the appended claims.

Claims

1. A method for predicting optical modem failure based on timing analysis and GRU network, characterized in that: Including steps: Obtaining an original training data set of the optical modem including basic operating status parameters, operating log data, and fault repair records, and preprocessing the original training data set to obtain a purely numerical operating status parameter data set and an operating log data set; Select an appropriate time window size and divide the operating state parameter data set and the operating log data set into time windows to obtain the total data set; specifically, select an appropriate time window size according to the granularity requirements. ; Divide the operating status parameter data set into time windows and convert the timestamp values into consecutive state parameter vectors into the same time window, and obtain the operation state parameter vector array of the corresponding time window; divide the operation log data set into time windows, and divide the log information whose occurrence time falls within a certain time window into the time window, and obtain the operation log vector array of the time window; obtain the total data set based on the operation state parameter vector array and the operation log vector array; Perform feature extraction of state parameter information and log information on each time window in the total data set to obtain a time window vector and a time window stream; specifically comprising: performing feature extraction on continuous value state parameter information, extracting one or more features including mean, variance, maximum value, minimum value and change trend, to obtain a first feature vector of each continuous value state parameter in the current time window; performing feature extraction on discrete value state parameter information, extracting one or more features including an initial stable value, number of changes and a final stable value, to obtain a second feature vector of each discrete value state parameter in the current time window; performing feature extraction on log information, sequentially traversing the configuration information sequence, alarm information sequence and error information sequence of each time window, counting the number of log information items of various configuration types, alarm types and error types, to obtain a third feature vector of the log information in the current time window; obtaining the time window vector and the time window stream based on the first feature vector, the second feature vector and the third feature vector; According to the fault repair record, a label value is marked for each time window; Taking the time window vector as input, building, training, and testing a GRU neural network model based on a training set and a test set divided by the time window stream according to a preset ratio to obtain a final-state GRU neural network model; An optical modem fault prediction value is obtained based on the final-state GRU neural network model, and an optical modem fault is warned based on the prediction value.

2. The optical modem fault prediction method according to claim 1, wherein: 。 3. The optical modem fault prediction method according to claim 1, wherein: Extracting the characteristics of the change trend specifically includes: Defining the monotonic coefficient , used to characterize the monotonicity of the continuous-valued state parameter within the current time window; Define the amplification factor , used to characterize the magnitude of the change in the continuous value state parameter within the current time window; Changing trends The calculation formula is: ； in, and is the weight coefficient, , , the weight coefficient is used to adjust and The relative weight of the impact on the changing trend.

4. The optical modem fault prediction method according to claim 1, wherein: The step of labeling each time window according to the fault repair record specifically includes: According to the fault repair record, all time points at which communication faults occurred are obtained, and the time window label value to which these time points belong is set to 1, indicating that the time window is the fault window in which the communication fault occurred; Defining constants , for the continuous time window, and set its label value to ,in , that is, the current window With the fault window The difference between the time points indicates the distance between the current window and the fault window; For all distances to the fault window The time windows are all considered normal windows and their label values are set to 0.

5. The optical modem fault prediction method according to claim 1, wherein: The constructed GRU neural network model specifically includes input layer, hidden layer and output layer; Input layer: The input parameter is the time window vector; Hidden layer: Use a single hidden layer structure; record the current input as , the reset gate and update gate are respectively denoted as and , the hidden state of the previous moment is recorded as , the hidden state at the current moment is recorded as , the candidate hidden state at the current moment is recorded as , then reset the gate and update gate The calculation formula is as follows: ， ， in, 、 、 and is the weight parameter, and is the bias parameter; Candidate hidden states The calculation formula is as follows: in, and is the weight parameter, is the bias parameter, symbol yes Product operator; use A nonlinear activation function is used to ensure that the value of the candidate hidden state remains in the interval (-1, 1). Its expression is: at last, Hidden state at the moment The calculation formula is as follows: Output layer: output predicted value , represents the probability of failure in the current operating state, and its expression is: in, is the output layer weight matrix, is the output layer bias parameter, using as an activation function.

6. The optical modem fault prediction method according to claim 5, wherein: Huber's RobustLoss is used as the loss function during training, and its expression is: in, Indicates the actual label value, Represents the predicted value of the model output; Use the mini-batch stochastic gradient descent algorithm to get the overall loss by taking the average error loss.

7. The optical modem fault prediction method according to claim 1, wherein: Training and testing the GRU neural network model based on the training set and test set divided according to a preset ratio of the time window stream specifically includes: The holdout method is used to divide the time window stream that has completed feature extraction into A continuous subsequence, and then The molecular sequences are divided into training set and test set according to the preset ratio, while keeping the proportion of subsequences containing fault windows in the training set and the test set equal to the total subsequences; The training set is input into the GRU neural network model, the learning rate is set to 0.03~0.1, the number of training iterations is 1000, and the model training is started; after the training is completed, the test set is input into the GRU neural network model, the test is performed to obtain the final loss value, and the learning rate with the minimum loss value is used to train the model again to obtain the final state GRU neural network model.

8. The optical modem fault prediction method according to claim 7, wherein: The obtaining of an optical modem fault prediction value based on the final-state GRU neural network model and issuing an early warning of the optical modem fault based on the prediction value specifically includes: The operating status parameters are read once every first preset time, and the operating log data is read once every second preset time. The status parameters and log data are merged, the time window features are extracted, and the time window is constructed. The time window vector is input into the final state GRU neural network model to obtain the output value. When the output prediction value is greater than the preset threshold, it is judged that a fault is about to occur and a fault alarm is issued. Otherwise, no response is given. At the same time, the hidden state of this time window is stored and participates in the hidden state calculation at the next time point, and this cycle is repeated.

Images